Red-Teaming Agent Execution Contexts: Open-World Security Evaluation on OpenClaw

A research team has released DeepTrap, an automated framework designed to uncover security flaws in agentic language-model systems by manipulating their execution contexts rather than user prompts, according to a paper submitted to arXiv on 11 May 2026 [1]. The framework targets OpenClaw, a system that relies on mutable elements such as files, memory, tools, and skills, which the authors argue create attack surfaces beyond direct user input [1]. DeepTrap treats adversarial context manipulation as a black-box trajectory-level optimization problem, balancing risk realization, benign-task preservation, and stealth [1]. It employs risk-conditioned evaluation, multi-objective trajectory scoring, reward-guided beam search, and reflection-based deep probing to identify compromised contexts that can trigger unsafe behavior while maintaining user-facing task completion [1]. The researchers constructed a 42-case benchmark spanning six vulnerability classes and seven operational scenarios [1]. Nine target models were evaluated using attack and utility grading scores [1]. The results indicate that final-response evaluation alone is insufficient for detecting such compromises, and the authors call for execution-centric security evaluation of agentic AI systems [1]. Agentic AI systems differ from narrow models in their ability to generalize knowledge and transfer skills across domains, a capability that companies such as OpenAI, Google, xAI, and Meta have stated as a goal [8]. The DeepTrap findings arrive as the AI industry undergoes rapid consolidation and valuation growth. Anthropic, a San Francisco-based AI safety company, reached an estimated valuation of $965 billion in May 2026, making it the most valuable pure-play AI company globally [6]. Google DeepMind, which merged with Google Brain in April 2023, has been responsible for developing the Gemini family of large language models and other generative tools [7]. The paper, titled "Red-Teaming Agent Execution Contexts: Open-World Security Evaluation on OpenClaw," was submitted by Hongwei Yao and revised on 14 June 2026 [1]. The initial submission was 3,603 KB, with the revised version reduced to 939 KB [1]. The code has been released on GitHub [2].